Can now query register blocksizes from blk algs.

Details:
- Added a new field to blksz_t objects that allows one to attach a
  sub-object. Doing this allows us to associate a register blocksize with
  any given cache blocksize. That way, the register blocksize can be
  queried wherever the cache blocksize would normally be accessible
  (e.g. a blocked algorithm).
- Modified bli_gemm_cntl.c (and 4m/3m variants) so that the register
  blocksizes are attached to the cache blocksizes after they are created.

frame/3/gemm/3m/bli_gemm3m_cntl.c

@@ -97,6 +97,12 @@ void bli_gemm3m_cntl_init()
 	                      BLIS_DEFAULT_3M_KR_Z, BLIS_EXTEND_3M_KR_Z );
 
 
+	// Attach the register blksz_t objects as sub-blocksizes to the cache
+	// blksz_t objects.
+	bli_blksz_obj_attach_to( gemm3m_mr, gemm3m_mc );
+	bli_blksz_obj_attach_to( gemm3m_nr, gemm3m_nc );
+	bli_blksz_obj_attach_to( gemm3m_kr, gemm3m_kc );
+
 
 	// Create function pointer object for each datatype-specific gemm
 	// micro-kernel.

frame/3/gemm/4m/bli_gemm4m_cntl.c

@@ -97,6 +97,12 @@ void bli_gemm4m_cntl_init()
 	                      BLIS_DEFAULT_4M_KR_Z, BLIS_EXTEND_4M_KR_Z );
 
 
+	// Attach the register blksz_t objects as sub-blocksizes to the cache
+	// blksz_t objects.
+	bli_blksz_obj_attach_to( gemm4m_mr, gemm4m_mc );
+	bli_blksz_obj_attach_to( gemm4m_nr, gemm4m_nc );
+	bli_blksz_obj_attach_to( gemm4m_kr, gemm4m_kc );
+
 
 	// Create function pointer object for each datatype-specific gemm
 	// micro-kernel.

frame/3/gemm/bli_gemm_cntl.c

@@ -97,6 +97,12 @@ void bli_gemm_cntl_init()
 	                      BLIS_DEFAULT_KR_Z, 0 );
 
 
+	// Attach the register blksz_t objects as sub-blocksizes to the cache
+	// blksz_t objects.
+	bli_blksz_obj_attach_to( gemm_mr, gemm_mc );
+	bli_blksz_obj_attach_to( gemm_nr, gemm_nc );
+	bli_blksz_obj_attach_to( gemm_kr, gemm_kc );
+
 
 	// Create function pointer object for each datatype-specific gemm
 	// micro-kernel.

frame/base/bli_blocksize.c

@@ -68,6 +68,16 @@ void bli_blksz_obj_init( blksz_t* b,
 	b->e[BLIS_BITVAL_DOUBLE_TYPE]   = be_d;
 	b->e[BLIS_BITVAL_SCOMPLEX_TYPE] = be_c;
 	b->e[BLIS_BITVAL_DCOMPLEX_TYPE] = be_z;
+
+	// By default, set the sub-blocksize field to NULL.
+	b->sub = NULL;
+}
+
+
+void bli_blksz_obj_attach_to( blksz_t* br,
+                              blksz_t* bc )
+{
+	bc->sub = br;
 }
 
 
@@ -119,6 +129,12 @@ dim_t bli_blksz_total_for_obj( obj_t*   obj,
 }
 
 
+blksz_t* bli_blksz_sub( blksz_t* b )
+{
+	return b->sub;
+}
+
+
 dim_t bli_determine_blocksize_f( dim_t    i,
                                  dim_t    dim,
                                  obj_t*   obj,
@@ -133,7 +149,7 @@ dim_t bli_determine_blocksize_f( dim_t    i,
 	// to bottom-right).
 
 	// Extract the execution datatype and use it to query the corresponding
-	// blocksize and blocksize extension values rom the blksz_t object.
+	// blocksize and blocksize extension values from the blksz_t object.
 	dt    = bli_obj_execution_datatype( *obj );
 	b_alg = bli_blksz_for_type( dt, b );
 	b_ext = bli_blksz_ext_for_type( dt, b );
@@ -173,7 +189,7 @@ dim_t bli_determine_blocksize_b( dim_t    i,
 	// to top-left).
 
 	// Extract the execution datatype and use it to query the corresponding
-	// blocksize and blocksize extension values rom the blksz_t object.
+	// blocksize and blocksize extension values from the blksz_t object.
 	dt    = bli_obj_execution_datatype( *obj );
 	b_alg = bli_blksz_for_type( dt, b );
 	b_ext = bli_blksz_ext_for_type( dt, b );
@@ -215,3 +231,20 @@ dim_t bli_determine_blocksize_b( dim_t    i,
 	return b_now;
 }
 
+
+dim_t bli_determine_reg_blocksize( obj_t*   obj,
+                                   blksz_t* b )
+{
+	num_t    dt;
+	blksz_t* b_sub_obj;
+	dim_t    b_sub;
+
+	// Extract the execution datatype and sub-blocksize and use them to
+	// query the the register blocksize from the blksz_t object.
+	dt        = bli_obj_execution_datatype( *obj );
+	b_sub_obj = bli_blksz_sub( b );
+	b_sub     = bli_blksz_for_type( dt, b_sub_obj );
+
+	return b_sub;
+}
+

frame/base/bli_blocksize.h

@@ -44,6 +44,9 @@ void bli_blksz_obj_init( blksz_t* b,
                          dim_t    b_c, dim_t be_c,
                          dim_t    b_z, dim_t be_z );
 
+void bli_blksz_obj_attach_to( blksz_t* br,
+                              blksz_t* bc );
+
 void bli_blksz_obj_free( blksz_t* b );
 
 dim_t bli_blksz_for_type( num_t    dt,
@@ -64,6 +67,8 @@ dim_t bli_blksz_ext_for_obj( obj_t*   obj,
 dim_t bli_blksz_total_for_obj( obj_t*   obj,
                                blksz_t* b );
 
+blksz_t* bli_blksz_sub( blksz_t* b );
+
 dim_t bli_determine_blocksize_f( dim_t    i,
                                  dim_t    dim,
                                  obj_t*   obj,
@@ -73,3 +78,5 @@ dim_t bli_determine_blocksize_b( dim_t    i,
                                  obj_t*   obj,
                                  blksz_t* b );
 
+dim_t bli_determine_reg_blocksize( obj_t*   obj,
+                                   blksz_t* b );

frame/include/bli_type_defs.h

@@ -416,10 +416,14 @@ typedef struct mem_s
 typedef struct blksz_s
 {
 	// Primary blocksize values.
-	dim_t v[BLIS_NUM_FP_TYPES];
+	dim_t           v[BLIS_NUM_FP_TYPES];
 
 	// Blocksize Extensions.
-	dim_t e[BLIS_NUM_FP_TYPES];
+	dim_t           e[BLIS_NUM_FP_TYPES];
+
+	// Sub-blocksize pointer.
+	struct blksz_s* sub;
+
 } blksz_t;
 
 // -- Function pointer object type --